A production-style Deep Learning NLP + Streamlit application that reads medical research abstracts and classifies each sentence into its functional role:

BACKGROUND • OBJECTIVE • METHODS • RESULTS • CONCLUSIONS

Unlike typical ML projects, this system:

• Trains the model from scratch inside the app
• Downloads and processes real-world data (~200K sentences)
• Provides a fully interactive UI for inference

• 🔥 End-to-end pipeline (data → training → inference → UI)
• 🧠 Advanced Tribrid Architecture (Token + Character + Positional)
• 📊 Real-world dataset (PubMed RCT)
• 🌐 Interactive Streamlit interface
• ⚡ No pre-trained local model required (self-training system)
• 🧩 Multi-input deep learning model (rare in beginner projects)

Medical abstracts are structured but unlabelled in raw form.

Example:

"This study evaluates..."
"We conducted a randomized trial..."
"Results showed improvement..."

➡️ Hard to scan quickly.

Automatically classify each sentence into:

• BACKGROUND
• OBJECTIVE
• METHODS
• RESULTS
• CONCLUSIONS

• 📚 PubMed 200K RCT Dataset
• 🔢 ~200,000 labeled sentences
• 🏷️ 5 structured classes

pubmed-rct/
└── PubMed_20k_RCT_numbers_replaced_with_at_sign/
    ├── train.txt   (~180K samples)
    ├── dev.txt     (~30K samples)
    └── test.txt

• TF-IDF + Logistic Regression

• 1D CNN
• LSTM
• Hybrid (Token + Character embeddings)

• ✅ Tribrid Model

  • Token embeddings (semantic understanding)
  • Character embeddings (morphological patterns)
  • Positional embeddings (structure awareness)

Input: Sentence
        │
        ├── Token Branch
        │     └── Universal Sentence Encoder (512-d)
        │     └── Dense Layer
        │
        ├── Character Branch
        │     └── TextVectorization
        │     └── Embedding (char-level)
        │     └── BiLSTM
        │
        └── Positional Branch
              ├── Line Number (one-hot)
              └── Total Lines (one-hot)

                     ↓
              Concatenation
                     ↓
              Dense → Dropout
                     ↓
              Softmax (5 classes)

• Loss: CategoricalCrossentropy (label_smoothing=0.2)
• Optimizer: Adam
• Output: Multi-class classification (5 labels)

When no model exists:

1. Clone dataset from GitHub (~177MB)
2. Parse and preprocess ~200K sentences
3. Encode labels (sklearn)
4. Build vectorization layers
5. Download Universal Sentence Encoder (~1GB, cached)
6. Train tribrid model
7. Save model → skimlit_tribrid_model.keras

⏱️ CPU Training Time: ~10–20 min per epoch

1. Paste medical abstract

2. Click Classify Sentences

3. Pipeline:

   • Sentence splitting (spaCy / fallback regex)
   • Multi-input preprocessing
   • Model prediction

4. Output:

   • Label per sentence
   • Confidence score
   • Clean UI display

project/
│
├── app.py                        # Streamlit app (training + inference)
├── SkimLit_1.ipynb              # Original model development
├── requirements.txt
├── README.md
├── sample_abstracts.txt
│
├── skimlit_tribrid_model.keras   # Generated after training
│
└── pubmed-rct/                   # Auto-downloaded dataset

git clone 
cd into project-folder

pip install -r requirements.txt

python -m spacy download en_core_web_sm

streamlit run app.py

• Dataset download
• Model training
• Model saved locally

• Model loads instantly
• No retraining

• ✅ Deep learning models outperform traditional ML
• ✅ Character embeddings improve robustness
• ✅ Positional embeddings add structural understanding
• ✅ Tribrid model achieves best performance

Available in:

sample_abstracts.txt

Includes real-world examples:

• Diabetes intervention
• COVID-19 vaccine study
• Hypertension trial
• Depression treatment
• Sleep study

• Python
• TensorFlow
• TensorFlow Hub
• Scikit-learn
• NumPy, Pandas
• Streamlit
• spaCy

• Python 3.10 / 3.11
• TensorFlow 2.15+
• Git installed